Methodology
How we turn messy, incompatible vendor metering into comparable costs — without pretending to a precision we don't have.
What we measure
AI vendors meter usage in incompatible units: input tokens, output tokens, cached tokens, reasoning tokens, credits, premium requests, subscription allocations, and raw compute time. TokenSpent normalizes these into one canonical usage model, then prices them with versioned, date-aware rates.
Separate views of cost
We keep two things apart on purpose: the money you were actually billed, and the reconstructed API-equivalent value of the same tokens. We never add them together — doing so would double count. Both are shown so you can reason about each.
Our trust rules
We hold ourselves to a short, strict list. We always:
- Label the provenance of every number.
- Show a confidence level, with the reason behind it.
- Preserve pricing history and use the rate that applied at the time.
- Show sample size before any statistical comparison.
- Distinguish official facts, user reports, estimates, and editorial judgement.
And we never:
- Fabricate users, customers, or project counts.
- Present sample data as real.
- Claim exact web-chat token usage without exact evidence.
- Use current pricing for historical projects silently.
- Treat unavailable data as zero.
- Rank a model as definitively “best” without evidence.